[DeepSparse Evaluation API] Perplexity

[dbogunowicz]

This PR adds the general support for perplexity integration evaluation
Supported and tested datasets to evaluate on openai_humaneval. The module is modular and expandable to support additional datasets

Example usage

Example using CLI:

(deepsparse_venv) damian@gpuserver6:/nm/drive0/damian/deepsparse$ deepsparse.eval hf:mgoin/TinyStories-1M-ds --dataset openai_humaneval --integration perplexity --limit 2 --batch_size 2

2024-02-05:14:08:00,089 INFO     [utils.py:148] Note: NumExpr detected 32 cores but "NUMEXPR_MAX_THREADS" not set, so enforcing safe limit of 8.
2024-02-05:14:08:00,089 INFO     [utils.py:160] NumExpr defaulting to 8 threads.
2024-02-05 14:08:07 deepsparse.evaluation.cli INFO     Creating deepsparse pipeline to evaluate from model path: hf:mgoin/TinyStories-1M-ds
2024-02-05 14:08:07 deepsparse.evaluation.cli INFO     Datasets to evaluate on: ['openai_humaneval']
Batch size: 2
Splits to evaluate on: None
Metrics to evaluate on: None
Additional integration arguments supplied: {'limit': 2}
DeepSparse, Copyright 2021-present / Neuralmagic, Inc. version: 1.7.0.20240104 COMMUNITY | (86c38139) (release) (optimized) (system=avx2, binary=avx2)
Fetching 11 files: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 11/11 [00:00<00:00, 25518.44it/s]
2024-02-05 14:08:10 deepsparse.evaluation.integrations.perplexity INFO     Argument `splits` is None. Defaulting to `test` split.██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4/4 [02:28<00:00, 37.16s/it]
2024-02-05 14:10:39 deepsparse.evaluation.cli INFO     Evaluation done. Results:
[Evaluation(task='perplexity', dataset=Dataset(type=None, name='openai_humaneval', config=None, split='test'), metrics=[Metric(name='perplexities', value=[7047.759765625, 11520.462890625, 7300.88671875, 6835.68017578125]), Metric(name='mean_perplexity', value=8176.197265625)], samples=None)]
2024-02-05:14:10:39,889 INFO     [cli.py:204] Evaluation done. Results:
[Evaluation(task='perplexity', dataset=Dataset(type=None, name='openai_humaneval', config=None, split='test'), metrics=[Metric(name='perplexities', value=[7047.759765625, 11520.462890625, 7300.88671875, 6835.68017578125]), Metric(name='mean_perplexity', value=8176.197265625)], samples=None)]
2024-02-05 14:10:39 deepsparse.evaluation.cli INFO     Saving the evaluation results to /nm/drive0/damian/deepsparse/result.json
2024-02-05:14:10:39,889 INFO     [cli.py:212] Saving the evaluation results to /nm/drive0/damian/deepsparse/result.json

Example using evaluate function:

from deepsparse import evaluate


result = evaluate(model="hf:mgoin/TinyStories-1M-ds",
         datasets="openai_humaneval", limit = 2, integration='perplexity')

print(result)

2024-02-05:14:00:38,576 INFO     [utils.py:148] Note: NumExpr detected 32 cores but "NUMEXPR_MAX_THREADS" not set, so enforcing safe limit of 8.
2024-02-05:14:00:38,576 INFO     [utils.py:160] NumExpr defaulting to 8 threads.
2024-02-05:14:00:45,891 WARNING  [__init__.py:194] Some tasks could not be loaded due to missing dependencies. Run with `--verbosity DEBUG` for full details.
DeepSparse, Copyright 2021-present / Neuralmagic, Inc. version: 1.7.0.20240104 COMMUNITY | (86c38139) (release) (optimized) (system=avx2, binary=avx2)
Fetching 11 files: 100%|██████████| 11/11 [00:00<00:00, 52015.04it/s]
2024-02-05:14:00:48,083 INFO     [perplexity.py:68] Argument `splits` is None. Defaulting to `test` split.
100%|██████████| 2/2 [01:43<00:00, 51.59s/it]
formatted=[Evaluation(task='perplexity', dataset=Dataset(type=None, name='openai_humaneval', config=None, split='test'), metrics=[Metric(name='perplexities', value=[7047.759765625, 11520.462890625]), Metric(name='mean_perplexity', value=9284.111328125)], samples=None)] raw=defaultdict(<class 'str'>, {'openai_humaneval': defaultdict(None, {'results': {'perplexities': array([ 7047.76 , 11520.463], dtype=float32), 'mean_perplexity': 9284.111}, 'split': 'test'})})

[bfineran]

can you add a plan in the in-code comment on how this can be abstracted out?

[dbogunowicz]

Just for bookkeeping, comment not relevant given the follow-up PR: #1586